Personalizing fitness recommendation systems is important to engage users of wearables and maximize their health outcomes. This research provides a new framework for the Internet of Things (IoT), which uses unsupervised clustering, collaborative filtering, and ensemble regression to make adaptive fitness recommendations. The methodology begins by grouping tracker activities and generating characteristics such as activity intensity and calorie expenditure per step. K-Means clustering helps to categorize users into five categories according to their activity level, from inactive to very active, for tailor-made recommendations. The lesson plans are adjusted according to the similarity in behavior using a collaborative filtering algorithm based on Nearest Neighbors (KNN) similarity. The random forest regression achieves near-perfect accuracy in predicting calorie burn (R2 = 0.9997), outperforming other regression models. In addition, the system simulates a real-time adjustable recommendation, recommending that users change the intensity of their exercise based on a predicted versus actual calorie deficit. Visual aids and the accuracy of the regression confirm the separation of clusters with different activity patterns. In contrast, qualitative analysis enhances the model’s ability to identify and distinguish different activity patterns. The results demonstrate the framework’s effectiveness in synthesizing user segmentation, predictive analytics, and dynamic recommendation and provide a solid basis for future innovations in real-time train adaptation.